Open MccreeZhao opened 5 years ago
To encourage spatial smoothness in the output image, we follow prior work on feature inversion and super -resolution, you can think of it as a regularization.
The differences between us and Justin Johnson's work in 2016 are the use of VGG19 instead of VGG16 and the HRNet.
By replacing the generation network with HRNet, our model attempts to assign different hue to different pixel boundaries, I think that's the reason why our output images more realistic.
Finally, I am busy looking for a teacher to pursue my master degree, so I may not have a lot of time to visit github. Any questions are advised to contact me by email: limingcv@gmail.com.
If you think this is a nice work, please give a star to the project, thanks! If you have a recommended tutor, please contact me, thanks!
It's a nice work! And after reading your paper, I wonder what is the total variation regularizer which you have mentioned in Sec3.2. Could you provide some explanations about that?
Besides, expect the Hi-Res Generation Network and the total variation regularizer loss, is there any difference between your work and Justin Johnson's work in 2016? What makes your output images more realistic? Thanks for your help